Skill ImproveSAFE

Turn Claude Code into a full game dev studio — 49 AI agents, 72 workflow skills, and a complete coordination system mirroring real studio hierarchy.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: skill-improve
description: "Improve a skill via a test-fix-retest loop — static checks, targeted fixes, keep or revert on score change."
argument-hint: "[skill-name]"
user-invocable: true
allowed-tools: Read, Glob, Grep, Write, Bash, Bash(bash "*/.claude/skills/skill-improve/../../hooks/yaml-helper.sh" resolve_config *)
model: sonnet
---

!`bash "${CLAUDE_SKILL_DIR}/../../hooks/yaml-helper.sh" resolve_config --keys automation`

**Automation mode**: Resolve `modes.automation` (`project.local.yaml` →
`project.yaml` → default `collaborative`). Every `AskUserQuestion` call and
every file write follows `.claude/docs/automation-modes.md`
(collaborative asks always · guided major-only · autonomous logs and proceeds;
`automation_always_ask` categories always prompt).

# Skill Improve

Runs an improvement loop on a single skill:
test → fix → retest → keep or revert.

---

## Phase 1: Parse Argument

Read the skill name from the first argument. If missing, output usage and stop:

```
Usage: /skill-improve [skill-name]
Example: /skill-improve tech-debt
```

Verify `.claude/skills/[name]/SKILL.md` exists. If not, stop with:
"Skill '[name]' not found."

---

## Phase 2: Baseline Test

Run `/skill-test static [name]` and record the baseline score:
- Count of FAILs
- Count of WARNs
- Which specific checks failed (Check 1–7)

Display to the user:
```
Static baseline:   [N] failures, [M] warnings
Failing: Check 4 (no ask-before-write), Check 5 (no handoff)
```

If the static result is **NOT ASSESSED** (the skill file could not be read or
parsed), stop and report it with its reason: there is no score to improve, and
0 FAILs from a file nobody could read is not a clean result.

If baseline is 0 FAILs and 0 WARNs, note it and proceed to Phase 2b.

### Phase 2b: Category Baseline

Look up the skill's `category:` field in `CCGS Skill Testing Framework/catalog.yaml`.

If no `category:` field is found, display:
"Category: not yet assigned — skipping category checks."
and skip to Phase 3.

If category is found, run `/skill-test category [name]` and record the category baseline:
- Count of FAILs
- Count of WARNs
- Which specific category rubric metrics failed

Display to the user:
```
Category baseline: [N] failures, [M] warnings  ([category] rubric)
```

A category result of **NOT ASSESSED** (a rubric section the skill-test could not
find or apply) is reported with its reason and is not a zero: the category
dimension is then excluded from "already passes" below.

If BOTH static and category baselines are 0 FAILs and 0 WARNs — and neither is
NOT ASSESSED — stop:
"This skill already passes all static and category checks. No improvements needed."

---

## Phase 3: Diagnose

Read the full skill file at `.claude/skills/[name]/SKILL.md`.

For each failing or warning **static** check, identify the exact gap:

- **Check 1 fail** → which frontmatter field is missing
- **Check 2 fail** → how many phases found vs. minimum required
- **Check 3 fail** → no verdict keywords anywhere in the skill body
- **Check 4 fail** → Write or Edit in allowed-tools but no ask-before-write language
- **Check 5 warn** → no follow-up or next-step section at the end
- **Check 6 warn** → `context: fork` set but fewer than 5 phases found
- **Check 7 warn** → argument-hint is empty or doesn't match documented modes

For each failing or warning **category** check (if category was assigned in Phase 2b),
identify the exact gap in the skill's text. For example:
- If G2 fails (gate mode, director panel width): skill body does not size the
  PHASE-GATE panel from `modes.workflow` (PR at `minimal`, TD + PR at `standard`,
  all 4 at `full`)
- If A2 fails (authoring, no per-section May-I-write): skill asks once at the end, not
  before each section write
- If T3 fails (team, BLOCKED not surfaced): skill doesn't halt dependent work on blocked agent

Show the full combined diagnosis to the user before proposing any changes.

---

## Phase 4: Propose Fix

Write a targeted fix for each failure and warning. Show the proposed changes
as clearly marked before/after blocks. Only change what is failing — do not
rewrite sections that are passing.

Ask: "May I write this improved version to `.claude/skills/[name]/SKILL.md`?"

If the user says no, stop here.

---

## Phase 5: Write and Retest

Record the current content of the skill file — the exact text, kept in this
conversation — so Phase 6 can restore it.

Write the improved skill to `.claude/skills/[name]/SKILL.md`.

Re-run `/skill-test static [name]` and record the new static score — measured by that run, never stated from the edit.
If a category was assigned, also re-run `/skill-test category [name]` and record the new category score.

Display the comparison:
```
Static:   Before [N] failures, [M] warnings  →  After [N'] failures, [M'] warnings
Category: Before [N] failures, [M] warnings  →  After [N'] failures, [M'] warnings  (if applicable)
Combined: [before] → [after] (improved / no change / worse)
```

`[before]` and `[after]` are the Phase 6 combined counts — failures plus
warnings, both dimensions — e.g. `Combined: 3 → 0 (improved)`.

---

## Phase 6: Verdict

Count the combined failure total: static FAILs + category FAILs + static WARNs + category WARNs.
**A re-test that comes back NOT ASSESSED** (for example, the edit broke the
frontmatter so the file no longer parses) is never an improvement, whatever its
counts: take the "did not improve" branch below.

**If combined score improved (combined failure count is lower than baseline):**
Report: "Score improved. Changes kept."
Show a summary of what was fixed in each dimension.

**If combined score is the same or worse:**
Report: "Combined score did not improve."
Show what changed and why it may not have helped.
Ask: "May I restore `.claude/skills/[name]/SKILL.md` to the content it had before this run?"
If yes: write the content recorded in Phase 5 back to the file with Write; if no, leave the file as written. Never
`git checkout` it — that retur

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.